# Copyright 2025 The EasyDeL Author @erfanzar (Erfan Zare Chavoshi).
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import typing as tp
from eformer.common_types import (
EMPTY,
MODE_TRAIN,
TP,
ColumnWise,
DynamicShardingAxes,
Replicated,
RowWise,
)
from easydel.infra.base_module import EasyDeLBaseConfig
from easydel.infra.etils import EasyDeLGradientCheckPointers
from easydel.infra.factory import register_config
from easydel.infra.utils import AttnMaskDetail, AttnMaskType
from easydel.layers.moe.utils import get_moe_partition_spec
[docs]class ExpertTensorParallel(DynamicShardingAxes):
"""Expert Tensor Parallelism (EPxTP) sharding axes."""
axes: tp.ClassVar = [TP, EMPTY, EMPTY]
mode: tp.ClassVar = MODE_TRAIN
[docs]@register_config("qwen2_moe")
class Qwen2MoeConfig(EasyDeLBaseConfig):
"""
Configuration objects inherit from [`EasyDeLBaseConfig`] and can be used to control the model outputs. Read
the documentation from [`EasyDeLBaseConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 151936):
Vocabulary size of the Qwen-2 MoE model. Defines the number of different tokens that can be represented by
the `inputs_ids` passed to the forward method.
hidden_size (`int`, *optional*, defaults to 2048):
Dimensionality of the encoder layers and the pooler layer.
intermediate_size (`int`, *optional*, defaults to 5632):
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
num_hidden_layers (`int`, *optional*, defaults to 24):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 16):
Number of attention heads for each attention layer in the Transformer encoder.
num_key_value_heads (`int`, *optional*, defaults to 16):
Number of key and value heads for each attention layer in the Transformer encoder.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) to use in the encoder and pooler. If string,
`"gelu"`, `"relu"`, `"swish"` and `"gelu_new"` are supported.
max_position_embeddings (`int`, *optional*, defaults to 32768):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 2048 or 4096).
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
rms_norm_eps (`float`, *optional*, defaults to 1e-6):
The epsilon used by the rms normalization layers.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether to tie the weights of the input embeddings and the output embeddings.
rope_theta (`float`, *optional*, defaults to 10000.0):
The theta value to use for rotary position embeddings.
use_sliding_window (`bool`, *optional*, defaults to `False`):
Whether to use a sliding window attention.
sliding_window (`int`, *optional*, defaults to 4096):
The sliding window size.
max_window_layers (`int`, *optional*, defaults to 28):
The maximum number of layers to use for the sliding window attention.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
decoder_sparse_step (`int`, *optional*, defaults to 1):
The sparse step for the decoder.
moe_intermediate_size (`int`, *optional*, defaults to 1408):
The intermediate size of the MoE layer.
shared_expert_intermediate_size (`int`, *optional*, defaults to 5632):
The intermediate size of the shared expert.
num_experts_per_tok (`int`, *optional*, defaults to 4):
The number of experts per token.
num_experts (`int`, *optional*, defaults to 60):
The number of experts.
norm_topk_prob (`bool`, *optional*, defaults to `False`):
Whether to normalize the top-k probabilities.
output_router_logits (`bool`, *optional*, defaults to `False`):
Whether to output the router logits.
router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
The coefficient for the router auxiliary loss.
mlp_only_layers (`list` of `int`, *optional*):
The layers that should only contain an MLP.
gradient_checkpointing (`str`, *optional*, defaults to `"nothing_saveable"`):
The gradient checkpointing configuration.
bits (`int`, *optional*):
The number of bits to quantize the model to.
"""
model_type: str = "qwen2_moe"
def __init__(
self,
vocab_size=151936,
hidden_size=2048,
intermediate_size=5632,
num_hidden_layers=24,
num_attention_heads=16,
num_key_value_heads=16,
hidden_act="silu",
max_position_embeddings=32768,
initializer_range=0.02,
rms_norm_eps=1e-6,
use_cache=True,
tie_word_embeddings=False,
qkv_bias=False,
rope_theta=10000.0,
use_sliding_window=False,
sliding_window=4096,
max_window_layers=28,
attention_dropout=0.0,
decoder_sparse_step=1,
moe_intermediate_size=1408,
shared_expert_intermediate_size=5632,
num_experts_per_tok=4,
num_experts=60,
norm_topk_prob=False,
output_router_logits=False,
router_aux_loss_coef=0.001,
mlp_only_layers=None,
gradient_checkpointing: EasyDeLGradientCheckPointers = EasyDeLGradientCheckPointers.NONE,
bits: int | None = None,
layer_types: list[str] | None = None,
**kwargs,
):
"""Initializes a Qwen2MoeConfig object.
Args:
vocab_size (int, optional): Vocabulary size. Defaults to 151936.
hidden_size (int, optional): Dimensionality of the embeddings and hidden states. Defaults to 2048.
intermediate_size (int, optional): Dimensionality of the intermediate layer in dense MLP. Defaults to 5632.
num_hidden_layers (int, optional): Number of hidden layers. Defaults to 24.
num_attention_heads (int, optional): Number of attention heads. Defaults to 16.
num_key_value_heads (int, optional): Number of key/value heads (for GQA). Defaults to 16.
hidden_act (str, optional): Activation function name. Defaults to "silu".
max_position_embeddings (int, optional): Maximum sequence length. Defaults to 32768.
initializer_range (float, optional): Standard deviation for weight initialization. Defaults to 0.02.
rms_norm_eps (float, optional): Epsilon for RMS normalization. Defaults to 1e-6.
use_cache (bool, optional): Whether to use KV cache. Defaults to True.
tie_word_embeddings (bool, optional): Whether to tie input/output embeddings. Defaults to False.
qkv_bias (bool, optional): Whether to include bias in QKV projections. Defaults to False.
rope_theta (float, optional): Base value for RoPE. Defaults to 10000.0.
use_sliding_window (bool, optional): Whether to use sliding window attention. Defaults to False.
sliding_window (int, optional): Sliding window size. Defaults to 4096.
max_window_layers (int, optional): Maximum number of layers for sliding window attention. Defaults to 28.
attention_dropout (float, optional): Dropout probability for attention scores. Defaults to 0.0.
decoder_sparse_step (int, optional): Frequency of MoE layers. Defaults to 1.
moe_intermediate_size (int, optional): Intermediate size for MoE MLP layers. Defaults to 1408.
shared_expert_intermediate_size (int, optional):
Intermediate size for the shared expert MLP. Defaults to 5632.
num_experts_per_tok (int, optional): Number of experts to route per token. Defaults to 4.
num_experts (int, optional): Total number of experts. Defaults to 60.
norm_topk_prob (bool, optional): Whether to normalize top-k probabilities in router. Defaults to False.
output_router_logits (bool, optional): Whether to output router logits. Defaults to False.
router_aux_loss_coef (float, optional): Coefficient for router auxiliary loss. Defaults to 0.001.
mlp_only_layers (list[int], optional):
List of layer indices that should only use MLP (no MoE). Defaults to None.
gradient_checkpointing (EasyDeLGradientCheckPointers, optional): Gradient checkpointing strategy.
Defaults to EasyDeLGradientCheckPointers.NONE.
bits (tp.Optional[int], optional): Quantization bits. Defaults to None.
**kwargs: Additional keyword arguments passed to the parent class.
"""
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.use_sliding_window = use_sliding_window
self.sliding_window = sliding_window
self.max_window_layers = max_window_layers
self.qkv_bias = qkv_bias
self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
self.use_cache = use_cache
self.rope_theta = rope_theta
self.attention_dropout = attention_dropout
# MoE arguments
self.decoder_sparse_step = decoder_sparse_step
self.moe_intermediate_size = moe_intermediate_size
self.shared_expert_intermediate_size = shared_expert_intermediate_size
self.num_experts_per_tok = num_experts_per_tok
self.num_experts = num_experts
self.norm_topk_prob = norm_topk_prob
self.output_router_logits = output_router_logits
self.router_aux_loss_coef = router_aux_loss_coef
self.gradient_checkpointing = gradient_checkpointing
self.bits = bits
self.mlp_only_layers = mlp_only_layers or []
self.layer_types = layer_types
if self.layer_types is None:
self.layer_types = [
"sliding_attention"
if self.sliding_window is not None and i >= self.max_window_layers
else "full_attention"
for i in range(self.num_hidden_layers)
]
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
[docs] def get_partition_rules(self, *args, **kwargs):
"""
Get the partition rules for the model.
Returns:
`tp.Tuple[tp.Tuple[str, PartitionSpec]]`: The partition rules.
"""
pmag = self.partition_manager
return (
(r"embed_tokens/embedding", pmag.resolve(ColumnWise)),
(r"self_attn/(q_proj|k_proj|v_proj)/kernel", pmag.resolve(ColumnWise)),
(r"self_attn/o_proj/kernel", pmag.resolve(RowWise)),
(r"self_attn/(q_proj|k_proj|v_proj)/bias", pmag.resolve(Replicated)),
(r"self_attn/o_proj/bias", pmag.resolve(Replicated)),
(r"mlp/(gate_proj|up_proj)/kernel", pmag.resolve(ColumnWise)),
(r"mlp/down_proj/kernel", pmag.resolve(RowWise)),
(r"mlp/.*proj/bias", pmag.resolve(Replicated)),
(r"mlp/gate/kernel", pmag.resolve(Replicated if self.use_expert_tensor_mode else ColumnWise)),
(r"mlp/gate/bias", pmag.resolve(Replicated)),
(
r"mlp/experts/(gate_proj|up_proj)/kernel",
get_moe_partition_spec(
partition_manager=self.partition_manager,
direction="column",
tensors_are_expert=self.use_expert_tensor_mode,
is_bias=False,
fsdp_is_ep_bound=self.fsdp_is_ep_bound,
sp_is_ep_bound=self.sp_is_ep_bound,
module_view=True,
),
),
(
r"mlp/experts/down_proj/kernel",
get_moe_partition_spec(
partition_manager=self.partition_manager,
direction="row",
tensors_are_expert=self.use_expert_tensor_mode,
is_bias=False,
fsdp_is_ep_bound=self.fsdp_is_ep_bound,
sp_is_ep_bound=self.sp_is_ep_bound,
module_view=True,
),
),
(r"mlp/experts/.*bias", pmag.resolve(Replicated)),
(r"mlp/shared_expert/(gate_proj|up_proj)/kernel", pmag.resolve(ColumnWise)),
(r"mlp/shared_expert/down_proj/kernel", pmag.resolve(RowWise)),
(r"mlp/shared_expert/.*bias", pmag.resolve(Replicated)),
(r"mlp/shared_expert_gate/kernel", pmag.resolve(ColumnWise)),
(r"mlp/shared_expert_gate/bias", pmag.resolve(Replicated)),
(r".*/(input_layernorm|post_attention_layernorm|norm)/kernel", pmag.resolve(Replicated)),
(r"lm_head/kernel", pmag.resolve(ColumnWise)),
(r"score/kernel", pmag.resolve(RowWise)),
(r".*bias", pmag.resolve(Replicated)),
(r".*", pmag.resolve(Replicated)),
)
@property
def granted_freq_max_position_embedding(self) -> int:
"""Returns the maximum position embedding size specifically for frequency-based position embeddings.
If `freq_max_position_embeddings` is set, it returns that value. Otherwise, it falls back to
`max_position_embeddings`.
Returns:
int: The granted maximum position embedding size for frequency encoding.
"""
return getattr(self, "freq_max_position_embeddings", self.max_position_embeddings)
@property
def granted_mask_max_position_embedding(self) -> int:
"""Returns the maximum position embedding size specifically for mask-based position embeddings.
If `mask_max_position_embeddings` is set, it returns that value. Otherwise, it falls back to
`max_position_embeddings`.
Returns:
int: The granted maximum position embedding size for mask encoding.
"""
return getattr(self, "mask_max_position_embeddings", self.max_position_embeddings)
[docs] def get_mask_details(self) -> dict[int, AttnMaskDetail]:
"""Retrieve attention mask details for each layer in the model.
This method generates a dictionary mapping layer indices to their corresponding attention mask details.
If a sliding window is defined, each layer is assigned a sliding window attention mask with the specified size.
Returns:
dict[int, AttnMaskDetail]: A dictionary where keys are layer indices (int) and values are AttnMaskDetail
objects specifying the attention mask type and size for each layer.
Notes:
- If `self.sliding_window` is None, an empty dictionary is returned.
- The method iterates over `self.num_hidden_layers` to assign mask details for each layer.
- The attention mask type is set to `AttnMaskType.SLIDING` when a sliding window is defined.
"""
mapping = {}
for layer_idx in range(self.num_hidden_layers):
if self.sliding_window is not None and self.use_sliding_window:
mapping[layer_idx] = AttnMaskDetail(mask_type=AttnMaskType.SLIDING, size=self.sliding_window)
return mapping